简介:AconicNewtonmethodisattractivebecauseitconvergestoalocalminimizzerrapidlyfromanysufficientlygoodinitialguess.However,itmaybeexpensivetosolvetheconicNewtonequationateachiterate.InthispaperweconsideraninexactconicNewtonmethod,whichsolvesthecouicNewtonequationoldyapproximatelyandinsonmunspecifiedmanner.Furthermore,weshowthatsuchmethodislocallyconvergentandcharacterizestheorderofconvergenceintermsoftherateofconvergenceoftherelativeresiduals.
简介:Thispaperrepresentsaninexactsequentialquadraticprogramming(SQP)algorithmwhichcansolvenonlinearprogramming(NLP)problems.Aninexactsolutionofthequadraticprogrammingsubproblemisdeterminedbyaprojectionandcontractionmethodsuchthatonlymatrix-vectorproductisrequired.SometruncatedcriteriaarechosensuchthatthealgorithmissuitabletolargescaleNLPproblem.Theglobalconvergenceofthealgorithmisproved.
简介:Inthispaperweconsidertheglobalconvergenceofanyconjugategradientmethodoftheformd1=-g1,dk+1=-gk+1+βkdk(k≥1)withanyβksatisfyingsumeconditions,andwiththestrongwolfelinesearchconditions.Undertheconvexassumptionontheobjectivefunction,weprevethedescenfpropertyandtheglobalconvergenceofthismethod.
简介:Inthisstudy,weuseinexactnewtonmethodstofindsolutionsofnonlinear,nondifferenti-ableoperatorequationsonBanachspaceswithaconvergencestructure.ThistechniqueinvolvestheintroductionofageneralizednormasanoperatorfromalinearspaceintoapartiallyorderedBanachspace.Inthiswaythemetricpropertiesoftheexaminedproblemcanbeanalyzedmoreprecisely.Moreover,thisapproachallmvsustoderivefromthesametheorem,ontheonehand,semi-localresultsofKantorovich-type,andontheotherhand,globalresultsbasedonmono-tonicityconsiderations.Furthermore,iveshowthatspecialcasesofourresultsreducetothecorrespondingonesalreadyintheliterature.Finally>ourresultsareusedtosolveintegralequationsthatcannotbesolvedwithexistingmethods.
简介:EllipticPDE-constrainedoptimalcontrolproblemswithL^1-controlcost(L^1-EOCP)areconsidered.TosolveL^1-EOCP,theprimal-dualactiveset(PDAS)method,whichisaspecialsemismoothNewton(SSN)method,usedtobeapriority.However,ingeneralsolvingNewtonequationsisexpensive.Motivatedbythesuccessofalternatingdirectionmethodofmultipliers(ADMM),weconsiderextendingtheADMMtoL^1-EOCP.TodiscretizeL^1-EOCP,thepiecewiselinearfiniteelement(FE)isconsidered.However,differentfromthefinitedimensionalL^1-norm,thediscretizedL^1-normdoesnothaveadecoupledform.Toovercomethisdifficulty,aneffectiveapproachisutilizingnodalquadratureformulastoapproximatelydiscretizetheL^1-normandL^2-norm.Itisprovedthattheseapproximationstepswillnotchangetheorderoferrorestimates.Tosolvethediscretizedproblem,aninexactheterogeneousADMM(ihADMM)isproposed.DifferentfromtheclassicalADMM,theihADMMadoptstwodifferentweightedinnerproductstodefinetheaugmentedLagrangianfunctionintwosubproblems,respectively.Benefitingfromsuchdifferentweightedtechniques,twosubproblemsofihADMMcanbeefficientlyimplemented.Furthermore,theoreticalresultsontheglobalconvergenceaswellastheiterationcomplexityresultso(1/k)forihADMMaregiven.Inordertoobtainmoreaccuratesolution,atwo-phasestrategyisalsopresented,inwhichtheprimal-dualactiveset(PDAS)methodisusedasapostprocessoroftheihADMM.Numericalresultsnotonlyconfirmerrorestimates,butalsoshowthattheihADMMandthetwo-phasestrategyarehighlyefficient.